GitHub Agentic Workflows reaches general availability
Markdown-defined coding agents can now run inside Actions on your own runners, with a model router, sandboxing and spend forecasting built in.
GitHub Agentic Workflows, the project that lets you write automation in Markdown and have a coding agent carry it out inside GitHub Actions, left preview and reached general availability on 28 October. It is built for the reasoning work a person usually does by hand: triaging issues, working out why a CI run failed, drafting documentation, writing tests.
The GA release adds Actions Runner Controller support with Docker-in-Docker, so workflows can run on your own self-hosted infrastructure instead of GitHub’s. You can now choose an engine and model provider per workflow, and a model router can send a task to whichever model fits its difficulty. Agents can also run inside a Cloud Hypervisor or NVX micro VM, and enclaves stop data leaking across repositories in workflows that touch more than one. The gh aw forecast command is no longer experimental, and now estimates both monthly spend and threat-detection credit use before anything runs. It’s available on GitHub Free, Team and Enterprise, though the default Copilot CLI engine needs a Copilot licence: bring your own agent if you don’t have one.
Why it matters: this is Actions plus an LLM, version-controlled like any other workflow file, with read-only permissions, sandboxed containers and threat-detection scanning on by default rather than something you bolt on yourself. If you’ve been wiring a coding agent into CI with scripts and good intentions, there is now a supported, GA path with the guardrails already built in.